A stochastic computing architecture for local contrast and mean image thresholding algorithm. (19th May 2022)

• Record Type: Journal Article
• Authors: Xu, Wenbing; Xie, Guangjun; Wang, Shaowei; Lin, Zhendong; Han, Jie; Zhang, Yongqiang
• Is Part Of: International journal of circuit theory and applications. Volume 50:Number 9(2022)
• Page Start: 3279
• Page End: 3291
• Publication Date: 2022-05-19
• Subjects: energy efficiency -- fault tolerance -- LCM algorithm -- stochastic computing; Electric circuit analysis -- Periodicals; 621.319205
• DOI: 10.1002/cta.3320
• Languages: English
• ISSNs: 0098-9886

Abstract

Abstract: Image binarization algorithms in document image analysis divide pixel values into two groups, including white as background and black as foreground. Among others, the local contrast and mean (LCM)‐based thresholding algorithm offers excellent performance in processing degraded documents. This algorithm, however, is susceptible to noise and requires significant hardware resources. In this paper, an energy‐efficient and fault‐tolerant architecture is proposed for implementing the LCM algorithm in stochastic computing (SC). Leveraging correlated input bitstreams, this architecture saves energy and improves the fault tolerance of the implementation. Experimental results show that the proposed LCM stochastic architecture significantly outperforms the stochastic implementation of the Sauvola algorithm in terms of both binarization accuracy and hardware overhead and energy consumption. Even using 16‐bit streams, the proposed circuit produces an error rate lower than 5%. The stochastic implementation of the LCM algorithm using a 16‐bit length FSM‐based LD sequence is 22 times less in area, 26 times less in total power, 28 times less in energy consumption and more fault‐tolerant than the conventional 8‐bit bit‐width weighted binary with the same frequency constraints. Abstract : The stochastic implementation of the LCM algorithm proposed in this paper has great advantages in terms of area, power consumption, and energy consumption compared to the conventional weighted binary implementation.